the bund conference | if there was a large model, would the “lehman brothers incident” not have happened?
2024-09-07
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
the bund summit showcases a number of key core technologies in the era of artificial intelligence.
according to the global digital economy white paper (2024), there are currently 30,000 ai core companies and more than 1,300 ai big models in the world. data from the ministry of industry and information technology shows that as of the end of july, there were 197 big models registered in china, of which 30% were general big models and 70% were industry big models.
when the spotlight of big models and generative artificial intelligence (aigc) comes on, people are more curious and complicated: when big models can chat, make videos, study finance and even create art, how broad are the boundaries of their imagination and creativity? at the same time, some people are worried that when aigc continuously "emerges" wisdom, where is the safety boundary?
expanding the boundaries of ai through technology optimization
during this year's bund conference, a number of new large models were unveiled, moving towards a more professional and segmented direction. the video large model vidu jointly released by shengshu technology and tsinghua university can generate high-definition videos with one click, igniting the spark of creativity; the intelligent microscopic imaging large model unifmir released by fudan university breaks through the limits of existing fluorescence microscopic imaging and greatly improves the analysis quality in fields such as life sciences...
there are also many large models in the field of financial technology. the "scholars and the world" financial model jointly built by tongji university and shanghai artificial intelligence laboratory made its debut at the bund conference. "through continued pre-training and task fine-tuning of financial corpus, it can improve the accuracy and speed of text recognition and analysis, and accelerate the process of information processing and decision-making." a member of the research team introduced that this financial model brings innovative experience and productivity improvement to the fields of intelligent decision-making and risk management.
big models are flourishing, but are there any limits to their creativity? sun maosong, foreign academician of the european academy of sciences and executive vice president of the institute of artificial intelligence at tsinghua university, gave his own judgment: "the boundaries of the creativity of big models are almost unlimited, but big models have strengths and weaknesses, which leads to the fact that their creativity in reality has limits."
in the view of yan junjie, founder and ceo of minimax, expanding the boundaries of ai requires patience, and he is currently more focused on expanding the boundaries of ai through technical optimization. he said that if the model error rate can be reduced to single digits, then ai may change from an assistant to a kind of productivity, bringing truly popular products and bringing essential changes to the industry. just a few days ago, minimax just released the video model abab-video-1, which can generate videos up to 6 seconds long by entering prompt words, but yan junjie is not in a hurry to move forward, "new things need to be debugged to a more satisfactory state before considering commercialization."
qiao yu, a leading scientist at the shanghai artificial intelligence laboratory, thinks about this question through the innovative evolution of multimodal large models. he believes that large models may take two development routes in the future. the first is the continuation of the large model route, which will generate stronger capabilities and penetrate into the industry by expanding the scale and more modes, but it will still have bottlenecks such as "hallucinations", "not knowing when you don't know", and will be accompanied by unsustainable problems such as extremely high energy consumption and extremely high costs; the second is the exploration of new technology routes, that is, exploring non-large model technology routes such as reinforcement learning, symbolic reasoning, and knowledge. he believes that the product that will eventually enable the large-scale application of large models must be a product that achieves a balance between the three core elements of professionalism, versatility, and economy.
above: the audience actually experienced that their thoughts and ideas blossomed into a flower through brain waves. below: children were attracted by the large models on display in the artificial intelligence laboratory.
the development of intelligent bodies needs to change from quantitative change to qualitative change
in the concrete manifestation of the creativity of large models, intelligent agent is a model that is relatively recognized by the industry. in the forum on "the creativity boundaries and application imagination of large models", "intelligent agent" is a hot word that appears frequently.
in sun maosong's view, the formation of intelligent agents requires four elements, namely the ability to generate large models from language, the ability to generate code from large models, the ability to generate large graphic models, and 3d modeling. "these elements have now formed relatively strong functions, making intelligent agents possible."
from the perspective of industry and industrial implementation, intelligent agents are actually downstream concepts of large models applied to various fields. sun maosong said that the commercial use of intelligent agents is worth looking forward to, but more patience is required. at present, many tasks can only be completed in a small part by intelligent agents driven by current large models, but as time goes by, when the capabilities of large models and intelligent agents are improved, the combination of the two can complete more tasks, which will also have greater commercial value.
at the bund conference, the world's leading companies brought more than 20 ai agents at once: digital avatars to assist in completing work, cyber pets to relieve loneliness, brain-computer aided deep sleep... in the "future wealth space", the new version of ant group's ai financial butler "ant xiaocai" was upgraded and unveiled. reporters found that when asking about the financial reports of listed companies, "ant xiaocai" can instantly generate a graphic version of the financial report interpretation, clearly presenting the changes in core data such as revenue. li yongsen, a professor at the institute of finance of the university of the chinese academy of social sciences, believes that in a rigorous field such as wealth management, this kind of attempt is relatively rare, but financial agents will appear sooner or later, which will help the high-end financial industry to shift from a product perspective to a user perspective.
sun maosong said that intelligent body technology is quietly entering people's lives, but this is a process of quantitative change leading to qualitative change, and it may take several years of iteration and accumulation before its performance will be significantly improved.
security is the bottom line in the financial technology field
when exploring the limits of imagination and creativity of big models, there is another limit that has also attracted sufficient attention from both inside and outside the industry - security. especially in the field of financial technology, security is the bottom line.
not long ago, the deepfake incident sounded the alarm. at the bund conference "aigc, release or weaken productivity" forum, reporters saw a real digital human "changing face" - his face changed into harry potter for a while, and then changed back to the appearance of the technology blogger "bi dao" himself, but it looked real and natural, even the digital human bi dao himself was frightened after watching it. "deepfake technology is a technical means of using deep learning methods to forge pictures, audio and video, making it possible to tamper with or even generate highly realistic and difficult to prove content. taking ai face-changing as an example, it can be easily used by criminals to commit financial fraud and cause property losses." bi dao said.
based on this, the bund conference launched a global deepfake attack and defense challenge. yao weibin, one of the contestants and technical director of zoloz, said that counterfeiting technology is constantly being updated and our ability to fight against it also needs to be continuously improved, which is the significance of launching the competition. more than a dozen institutions also jointly released the first domestic "ai face-changing" detection standard for financial scenarios, providing a basis for the security detection and evaluation of fake digital faces in financial scenarios, and filling the gap in this field.
starting from the "damage value" of ai face-changing to financial security, jiang changjun, an academician of the chinese academy of engineering, is thinking about the feasibility of large financial models and large devices in the field of financial systemic security: in 2008, the lehman brothers incident caused the united states to take a big detour in the development of the financial derivatives market. it is very secretive, but now using the computing power of big data, can we effectively track and prevent such incidents from happening?
jiang changjun believes that the large-scale scientific equipment of digital finance has a unique role and advantages in the field of financial risk prevention, and shanghai has the greatest development potential in this regard. "as an international financial center, how shanghai can take another step forward and improve itself, inject these scientific and technological elements and scientific and technological forces into large models and large equipment, and better prevent systemic financial risks, is a proposition worth thinking about."
author: xu jinghui
text: our reporter xu jinghui photo: oriental ic/xu jinghui editor: shi wei
please indicate the source when reprinting this article.
